String tensor support (#90)

* added support for string tensors,
when the data is given by VALUE (and not with BLOB)

* updated test.py
for setting and getting string tensors by VALUE

* added support for tensorset
from numpy string array as blob

* updated test.py
to test tensorset with numpy string array

* linting

* small fix

* Review fixes:

Added a comment.

Deleted numpy_string2blob and replaced with a single line using join.

Deleted utils.recursive_bytetransform_str and
sets 'target' to a decode function

Co-authored-by: alonre24 <[email protected]>

redisai/command_builder.py

@@ -176,7 +176,12 @@ def tensorset(
         args = ["AI.TENSORSET", key, dtype, *shape, "BLOB", blob]
     elif isinstance(tensor, (list, tuple)):
         try:
-            dtype = utils.dtype_dict[dtype.lower()]
+            # Numpy 'str' dtype has many different names regarding maximal length in the tensor and more,
+            # but the all share the 'num' attribute. This is a way to check if a dtype is a kind of string.
+            if np.dtype(dtype).num == np.dtype("str").num:
+                dtype = utils.dtype_dict["str"]
+            else:
+                dtype = utils.dtype_dict[dtype.lower()]
         except KeyError:
             raise TypeError(
                 f"``{dtype}`` is not supported by RedisAI. Currently "

redisai/postprocessor.py

@@ -42,7 +42,11 @@ def tensorget(res, as_numpy, as_numpy_mutable, meta_only):
                 mutable=False,
             )
         else:
-            target = float if rai_result["dtype"] in ("FLOAT", "DOUBLE") else int
+            if rai_result["dtype"] == "STRING":
+                def target(b):
+                    return b.decode()
+            else:
+                target = float if rai_result["dtype"] in ("FLOAT", "DOUBLE") else int
             utils.recursive_bytetransform(rai_result["values"], target)
             return rai_result
 

redisai/utils.py

@@ -15,6 +15,7 @@
     "uint32": "UINT32",
     "uint64": "UINT64",
     "bool": "BOOL",
+    "str": "STRING",
 }
 
 allowed_devices = {"CPU", "GPU"}
@@ -24,11 +25,15 @@
 def numpy2blob(tensor: np.ndarray) -> tuple:
     """Convert the numpy input from user to `Tensor`."""
     try:
-        dtype = dtype_dict[str(tensor.dtype)]
+        if tensor.dtype.num == np.dtype("str").num:
+            dtype = dtype_dict["str"]
+            blob = "".join([string + "\0" for string in tensor.flat])
+        else:
+            dtype = dtype_dict[str(tensor.dtype)]
+            blob = tensor.tobytes()
     except KeyError:
         raise TypeError(f"RedisAI doesn't support tensors of type {tensor.dtype}")
     shape = tensor.shape
-    blob = bytes(tensor.data)
     return dtype, shape, blob
 
 
@@ -38,7 +43,9 @@ def blob2numpy(
     """Convert `BLOB` result from RedisAI to `np.ndarray`."""
     mm = {"FLOAT": "float32", "DOUBLE": "float64"}
     dtype = mm.get(dtype, dtype.lower())
-    if mutable:
+    if dtype == 'string':
+        a = np.array(value.decode().split('\0')[:-1], dtype='str')
+    elif mutable:
         a = np.fromstring(value, dtype=dtype)
     else:
         a = np.frombuffer(value, dtype=dtype)

test/test.py

@@ -117,6 +117,12 @@ def test_set_non_numpy_tensor(self):
         self.assertEqual([2, 2], result["shape"])
         self.assertEqual("BOOL", result["dtype"])
 
+        con.tensorset("x", (12, 'a', 'G', 'four'), dtype="str", shape=(2, 2))
+        result = con.tensorget("x", as_numpy=False)
+        self.assertEqual(['12', 'a', 'G', 'four'], result["values"])
+        self.assertEqual([2, 2], result["shape"])
+        self.assertEqual("STRING", result["dtype"])
+
         with self.assertRaises(TypeError):
             con.tensorset("x", (2, 3, 4, 5), dtype="wrongtype", shape=(2, 2))
         con.tensorset("x", (2, 3, 4, 5), dtype="int8", shape=(2, 2))
@@ -156,6 +162,12 @@ def test_numpy_tensor(self):
         self.assertEqual(values.dtype, "bool")
         self.assertTrue(np.array_equal(values, [True, False]))
 
+        input_array = np.array(["a", "bb", "⚓⚓⚓", "d♻d♻"]).reshape((2, 2))
+        con.tensorset("x", input_array)
+        values = con.tensorget("x")
+        self.assertEqual(values.dtype.num, np.dtype("str").num)
+        self.assertTrue(np.array_equal(values, [['a', 'bb'], ["⚓⚓⚓", "d♻d♻"]]))
+
         input_array = np.array([2, 3])
         con.tensorset("x", input_array)
         values = con.tensorget("x")
@@ -174,10 +186,6 @@ def test_numpy_tensor(self):
         np.put(ret, 0, 1)
         self.assertEqual(ret[0], 1)
 
-        stringarr = np.array("dummy")
-        with self.assertRaises(TypeError):
-            con.tensorset("trying", stringarr)
-
     # AI.MODELSET is deprecated by AI.MODELSTORE.
     def test_deprecated_modelset(self):
         model_path = os.path.join(MODEL_DIR, "graph.pb")

tox.ini

@@ -6,7 +6,7 @@ envlist = linters,tests
 max-complexity = 10
 ignore = E501,C901
 srcdir = ./redisai
-exclude =.git,.tox,dist,doc,*/__pycache__/*
+exclude =.git,.tox,dist,doc,*/__pycache__/*,venv
 
 [testenv:tests]
 whitelist_externals = find